Improving Breast Cancer Detection and Diagnosis with CAD
Computer-aided detection (CAD) of breast cancer is rapidly becoming a well accepted clinical practice. Studies have found that radiologists'attitude toward and acceptance of CAD-cued micro-calcification clusters and masses were substantially different. Due to the high sensitivity, radiologists heavily rely on CAD-cued results while searching for micro-calcifications. However, the lower CAD sensitivity for mass detection (including a fraction of subtle masses being cued only on one view) and the higher false-positive detection (FP) rates reduce radiologists'confidence in CAD-cued masses. As a result, radiologists frequently discard CAD-cued subtle masses in the clinical practice. To improve CAD performance and increase radiologists'confidence in using CAD-cued masses in their decision making, we propose two observer-focused innovative approaches to develop and optimize CAD schemes. By maintaining a comparable FP rate to current commercial CAD systems, the new approaches aim to either increase the number of masses being cued on both ipsilateral (CC and MLO) views or cue more subtle masses by eliminating a fraction of other regions that can be easily identified and classified by radiologists without using CAD. To test these approaches, we propose three specific tasks. First, we will develop a unique multi-view based CAD scheme. To more sensitively detect and better match subtle mass regions, we introduce a concept of limited viewing of specific regions into the arena of CAD development. After detecting a matching strip on the ipsolateral view, the scheme applies a second highly sensitive detection scheme only to this strip to identify matched regions. To control for and reduce FP rates, the scheme limits the number of possible matched candidates to less than one per image. Second, we will develop an integrated CAD scheme that includes a combined score for both detection and classification. To improve direct use of features computed by the detection module in the classification task, we will apply a new dual active contour algorithm that should improve mass region segmentation. We will separately optimize two machine learning classifiers to generate a detection score (the likelihood of being a true-positive mass) and a classification score (the likelihood of each detected mass for malignancy) for each segmented region. We will then develop a fusion method to combine these two scores and generate a new summary index that is more heavily weighted for subtle masses. Using this scheme, we can change the current mass detection based cuing method to a new cancer-based cueing method. Third, we will conduct a pilot observer performance study to investigate radiologists'performance under three CADcueing modes (using the current commercial single-image based, the new multi-view based, and the new integrated CAD schemes). The reading results will be compared and analyzed using both ROC and JAFROC methodologies. We note that the approach is substantially different than focusing on incremental improvements in image based detection schemes in that the observer's actual use (or not) of the CADcued regions drives our objectives in this project, resulting in a targeted development effort.